Re: 5.10 kernel series fails to build on newer toolchain: FAILED unresolved symbol filp_close

["David Nyström" <david.nystrom@est.tech>]

[-- Attachment #1: Type: text/plain, Size: 1793 bytes --]


On Sun, 26 Oct 2025, Alyssa Ross wrote:

> On Tue, Jun 24, 2025 at 09:05:31AM +0200, Philip Müller wrote:
>> On 6/5/25 10:46, Greg KH wrote:
>>> I have no context here, sorry...
>>
>> Seems with 5.10.239-rc1 it compiles again just fine ...
>
> We've been seeing this issue[1] since 5.10.244 (specifically commit
> b039655d31a1 ("genirq: Provide new interfaces for affinity hints")),
> and still in 5.10.245.
>
> Given that this has apparently come up before, and I don't see any
> likely cause looking at that diff, I suppose it's probably some build
> issue in 5.10 that can be triggered by innocent diffs, and so is liable
> to keep fixing and breaking itself until somebody figure out the root
> cause…
>
> [1]: https://github.com/NixOS/nixpkgs/pull/448034#issuecomment-3364278085
>

I have analyzed the issue, and it seems like never versions of GCC will 
agressively inline both in compile and linktime.
I'm not sure if this is fixed properly in master, I would assume its fixed 
by adding enough code into BTF exposed functions so that they not
optimized away anymore.

I can "fix" this locally via:
CONFIG_DEBUG_INFO_BTF=n

or

diff --git a/fs/Makefile b/fs/Makefile
index c660ce28f149..6fbaccbae1d9 100644
--- a/fs/Makefile
+++ b/fs/Makefile
@@ -6,6 +6,10 @@
  # Rewritten to use lists instead of if-statements.
  #

+# Prevent agressive link-time inlining of BTF symbols
+# Prevent GCC 15 IPA from removing filp_close symbol needed for BTF
+CFLAGS_open.o = -fno-inline -fno-ipa-sra
+

or by removing line below.

linux-stable$ git grep filp_close kernel/trace
kernel/trace/bpf_trace.c:BTF_ID(func, filp_close)

Last seems like the simplest option, but would cause regressions for BPF
programs I presume.
GCC 15+ does not seem to adhere to __noinline, at least not for this 
issue.